Abstract
Large-volume 3D dense prediction is essential in industrial applications like energy exploration and medical image segmentation. However, existing deep learning models struggle to process full-size volumetric inputs at inference due to memory constraints and inefficient operator execution. Conventional solutions-such as tiling or compression-often introduce artifacts, compromise spatial consistency, or require retraining. Here we present a retraining-free inference optimization framework that enables accurate, efficient, whole-volume prediction without performance degradation. Our approach integrates operator spatial tiling, operator fusion, normalization statistic aggregation, and on-demand feature recomputation to reduce memory usage and accelerate runtime. Validated across multiple seismic exploration models, our framework supports full size inference on volumes exceeding 1024(3) voxels. On FaultSeg3D, for instance, it completes inference on a 1024(3) volume in 7.5 seconds using just 27.6 GB of memory-compared to conventional inference, which can handle only 448(3) inputs under the same budget, marking a 13 × increase in volume size without loss in performance. Unlike traditional patch-wise inference, our method preserves global structural coherence, making it particularly suited for tasks inherently incompatible with chunked processing, such as implicit geological structure estimation. This work offers a generalizable, engineering-friendly solution for deploying 3D models at scale across industrial domains.